Research And Methods Of Identifying Influential Nodes In Complex Networks Based On Deep Learning

Complex network is a highly simplified model of various complex systems in the real world,and it has been widely used in many fields,such as social networks,biological networks,etc.The study found that for a large-scale complex network,a small number of influential nodes play a crucial role in information dissemination,which are called key nodes.The key node identification problem in a complex network is mainly divided into two categories:one is the influence maximization problem;the other is the cascading failure caused by destroying the network structure.Among them,the main content discussed in this thesis is the influence maximization problem.At present,many methods have been proposed to identify single key node in complex networks,ranging from node centrality to dynamic principle.However,most of them only consider the network structure or node characteristics to evaluate the importance of nodes.Considering the above problems,this thesis transforms the key node identification problem in complex network into a classification problem based on graph neural network.Combined with graph neural network and attention mechanism,and based on the method of node similarity,the weight of neighbor nodes is calculated,which solves the problem that the importance of neighbor nodes cannot be distinguished.A more reasonable graph neural network KAGNN is proposed to solve the problem of node importance ranking.But in most real cases,researchers pay more attention to how to select the key node group for maximum influence.Because the rich club phenomenon in the complex network needs to be considered,rather than simply choosing the nodes with the highest influence by ranking results.Many scholars have proposed some greedy algorithms and heuristic algorithms,and even introduced deep learning models to find the optimal solution of node groups.Therefore,this thesis also focuses on the problem of maximizing the influence of key node groups.Considering the information spread among nodes and the phenomenon of rich clubs,this thesis proposes a reordering algorithm based on community detection based on the information spread function.For each iteration of the proposed algorithm,the proposed algorithm selected node with the highest scores,and then with the help of the information transmission probability function,the selected node of the neighbor nodes selection algorithm for the qualified nodes and update their records.On datasets of multiple synthetic networks and real networks,a large number of experiments based on the SIR propagation model show that the performance of KAGNN in identifying key nodes is better than that of traditional benchmark methods,and the CINF re-ranking algorithm compared with the original methods has increased from 2.6% to 29.2%,the best performing re-ranking method has a certain improvement in performance on each network compared with the key node group benchmark method.Through experimental analysis,the KAGNN model and the CINF algorithm can effectively improve the network propagation scale and the average distance between node groups,especially when the CINF algorithm is combined with various node importance ranking algorithms,the propagation scale of the node set in the SIR propagation model can be greatly improved.